marchgad / csdb

A Deep Backend for Speaker Verification Based on Xvector Coupling Space and Metric Learning.

CSDB

A Deep Backend for Speaker Verification Based on Xvector Coupling Space and Metric Learning.

Requirements

• python3

• torch

• kaldiio

• kaldi

Others

LDA and PLDA generated from kaldi, zero centering is unnecessary for LDA.

Results

Models of results we showed in paper didn't trained really well at that time, so the following results may be better.

Name	Extractor	Network	GPLDA EER(%)	CSDB EER(%)
FC A	TDNN	kaldi_lda	3.1177	2.4549
CNN	TDNN	kaldi_cnn	3.1177	2.7678
FC B	ResNet	kaldi_lda	2.3171	1.9459
FC C	EFTDNN	kaldi_lda	1.0392	0.9067
NPLDA	TDNN	NPLDA/LDC	3.1177	2.5875

Triplet demo

This section is for our best performance triplet network. A light demo with needed data has been prepared, please run the demo to check the environment is fully setted.

Note that all the following commands are started under './CSDB/.'

step 1: use absolute path in '.json' and '.scp'

python utils/scp_relative2abs.py scps/demo
python utils/json_relative2abs.py json_configs/demo.json

step 2: run demo

python train.py json_configs/demo.py &

step 3: see the log

Log is in 'exp/demo/log'

tail -f exp/demo/log

step 4: test

cd exp/demo
cat result

or we can score by ourselves.

cd exp/demo/score
python get_scores.py -e 1
// wait for scoring
cat result

see validate result

cat exp/demo/nnet/validate_eer_mindcf

Siamese

Training for siamese network is slightly different from triplet in data selection. Thus, we provide another script 'LDC_train.py' for siamese training such as NPLDA or LDC/HDC in our CSDB.

To run siamese and better compare with NPLDA, we follow its way on data preparing which generates a tsv file for training. After that, we replace the 'train.py' in step 2 with 'LDC_train.py'.